| Title |
Real-world Face Super-resolution via Cyclic Network and Domain Adversarial Loss using Unpaired Data |
| Authors |
최익규(Ikkyu Choi) ; 박해솔(Haesol Park) ; 김익재(Ig-Jae Kim) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.8.35 |
| Keywords |
AI; Deep learning; Face super-resolution; Real-world face super-resolution; Unpaired dataset |
| Abstract |
Most deep-learning-based face super-resolution networks has been trained under the assumption that they know how input low-resolution images are generated from original high-resolution facial images (i.e. bicubic downsampling). For this reason, it is difficult to obtain clear super-resolution results as reported in papers for the real-world low-resolution face image that was not seen during training. In order to overcome this limitation, we propose a novel method that can roboustly perform face super-resolution on real-world low-resolution faces by utilizing actual low-resolution images(unpaired) alo-ng with commonly used dataset(paired). Specifically, we designed a cycle-based network structure so that a low-resolution face image without a corresponding high-resolution image could be used for super-resolution training. In addition, we pro-pose a method to apply domain adversarial loss to super-resolution research, which helps the proposed network perform s-table learning even though it trained from a dataset composed of two different domains (paired and unpaired). Through t-he several experiments, our method outperforms state-of-the-art methods in both qualitatively and quantitatively metrics at not only synthetic low-resolution face images but also real-world low-resolution face images. |